NVIDIA Kepler speculation thread

[AnarchX]

http://www.expreview.com/18768-4.html

 

[fellix]

Huh?

 

[Man from Atlantis]

http://www.expreview.com/18768-4.html

thats pretty much confirms what charlie says 8 and 7 parts with 2 GPCs 8SMs in total.. so GK110 4GPCs and 128CCs per SMs x16SMs, 2048CCs total, 128TMUs, 384 Bit 48 ROPs.. i expect pixel export doubles with new SMs..

Spoiler

 

[chavvdarrr]

Huh?

384=?

Spoiler

фърляй око в хвбг, да не се чудиш има ли хотклок няма ли

 

[fellix]

Does that mean the SIMD width in Kepler will now match the warp size -- 32 lanes?

That would make for 6 SIMDs per multiprocessor. I wonder how the scheduling and RF access are designed around this...

 

[Picao84]

thats pretty much confirms what charlie says 8 and 7 parts with 2 GPCs 8SMs in total.. so GK110 4GPCs and 128CCs per SMs x16SMs, 2048CCs total, 128TMUs, 384 Bit 48 ROPs.. i expect pixel export doubles with new SMs..

Spoiler

Is it my impression or the card in the powerpoint does not match the card we have been seeing around? The position of the fan does look diferent to me, unless the card in the slide is in an inverted position, with the PCI-E connection up.

 

[lanek]

192cuda cores / SM ( 8 SM ) ?

 

[AnarchX]

Planets fastest one was GTX 580: http://www.tomshardware.com/news/gtx-580-vapor-chamber-heatsink-cooler,11597.html
The texts asks if they can reclaim the title.

GK104 should be able to keep up with GF110 in geometry/tessellation, even if they put the units in a more compact order.

 

[Picao84]

Planets fastest one was GTX 580: http://www.tomshardware.com/news/gtx-580-vapor-chamber-heatsink-cooler,11597.html.

Ok, that was a GTX580 slide. I knew something was not fitting

 

[Man from Atlantis]

btw NDA is lifting at 6 AM PST, 22th..
TMU count is still unkown, might be 64TAU+64TFU, 64TAU+128TFU or 128TAU+128TFU..

 

[UniversalTruth]

With regards to the first part, it doesn't matter what Apple thinks if Nvidia can't supply enough chips. What would you propose they do? Cut Macbook shipments in half (or whatever number) instead? It isn't as if those lower priced Macbooks won't sell regardless of what graphics drives them. This, of course, presumes that the rumors of Nvidia not being able to provide enough Keplar chips is true.

With regards to the second part. Well yes, Nvidia is certainly hoping that HPC and professional sales continue to grow. As well they do have their Tegra initiative for low range and potentially midrange graphics in the future.

Windows 8 might allow Tegra valuable inroads into the budget desktop PC market they've been squeezed out of. It becomes even more important if developers bother to port AAA budgeted games to Metro (and hence capability of running on Arm based computers).

And as long as PC gaming continues to exist, there will always be demand for enthusiast and performance discrete graphics cards.

Regards,
SB

Going for AMD APUs (here I hope the situation in GloFo is not that bad) or discrete graphics chips (here I hope AMD don't have the same supply troubles).

Given how late it is in the game, and how long it takes to retool a laptop’s physical parts, we don’t think Apple will deviate from the current plan at the higher end of the lineup. We also doubt that AMD will get any business out of it, but you never know. Until the complete stoppage at TSMC, AMD was not having yield issues like Nvidia, but that doesn’t mean wafer supply was enough to supply their existing demands, much less take on Apple as a customer at the last minute.

http://semiaccurate.com/2012/03/13/apple-drops-nvidia-kepler-from-large-numbers-of-laptops/

 

[MDolenc]

btw NDA is lifting at 6 AM PST, 22th..
TMU count is still unkown, might be 64TAU+64TFU, 64TAU+128TFU or 128TAU+128TFU..

How about 96TAU+96TFU?

 

[itaru]

multipurpose arithmetic functional unit ?

 

[Love_In_Rio]

How about 96TAU+96TFU?

Special units for deferred engines like Battlefield 3 or for new SRAA ?.

 

[OlegSH]

Special units for deferred engines like Battlefield 3 or for new SRAA ?.

TAU = texture adress unit
TFU = texture filtering unit

 

[Gipsel]

btw NDA is lifting at 6 AM PST, 22th..
TMU count is still unkown, might be 64TAU+64TFU, 64TAU+128TFU or 128TAU+128TFU..

Anything less than 128 filter units would surprise me. 16 SMs each containing 96SPs (3x32) with 8 TMUs it's going to be.

 

[iMacmatician]

thats pretty much confirms what charlie says 8 and 7 parts with 2 GPCs 8SMs in total.. so GK110 4GPCs and 128CCs per SMs x16SMs, 2048CCs total, 128TMUs, 384 Bit 48 ROPs..

Is it possible to have a 12 SM x 192 CC/SM for GK110?

 

[fellix]

 

[m4trix]

Then what does it use? .....

 

[Man from Atlantis]

well, considering it's fair bit smaller than GF104/114; 4xCCs, no SFUs, it's reasonable they keep TAU count same and just double TFUs(64/128).. and 30% clock bump should be enough for TAUs.. or what MDolenc says is also possible 12 Fermi TMUs per SM..

but given Nvidia's earlier track record one arch they keep TAU/TFU at same count and on next arch doubles the TFUs while keeping TAUs same, i ask if with Kepler they continue the trend

Spoiler

Is it possible to have a 12 SM x 192 CC/SM for GK110?

i find it unlikely they reduce GPC count from Fermi unless they decide to use just 2 SMs per GPC..